-
Notifications
You must be signed in to change notification settings - Fork 18
/
roi_nn.py
executable file
·202 lines (144 loc) · 6.81 KB
/
roi_nn.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Tue Aug 4 19:55:40 2020
@author: asabater
"""
import os
from emodel import darknet_body
from keras.layers import Input, Dense, Flatten, concatenate, UpSampling2D, Conv2DTranspose, Lambda, GlobalAveragePooling2D, Reshape
from keras.layers.normalization import BatchNormalization
from keras.engine.topology import Layer
from keras.models import Model
from keras.layers import Input, Reshape
from keras.models import model_from_json
import numpy as np
from ROI_pooling_frcnn import RoiPoolingConv
from keras import backend as K
from tqdm import tqdm
import json
# Creates and initializes the YOLOv3 backbone for the feature maps extraction
def get_backbone(path_weights=None, downsample_rate=32):
dummy_img = Input(shape=(None,None,3), name='dummy_img')
backbone = darknet_body(dummy_img)
backbone = Model(dummy_img, backbone, name='backbone')
if downsample_rate == 32: backbone = backbone.layers[:] # 13x1054
elif downsample_rate == 16: backbone = backbone.layers[:153] # 26x512
elif downsample_rate == 8: backbone = backbone.layers[:93] # 52x256
elif downsample_rate == 4: backbone = backbone.layers[:33] # 104x128
elif downsample_rate == 2: backbone = backbone.layers[:15] # 208x64
elif downsample_rate == 1: backbone = backbone.layers[:4] # 416x32
elif downsample_rate == 0: backbone = backbone.layers[:1] # 413x3
else: raise ValueError('donwsample_rate not valid')
backbone = Model(dummy_img, backbone[-1].output, name='backbone')
print('Last layer {}: {}'.format(len(backbone.layers), backbone.layers[-1]))
if path_weights is not None:
print('** Pretraining backbone **')
backbone.load_weights(path_weights, by_name=True, skip_mismatch=False)
else:
print('** No pretraining **')
for i in range(len(backbone.layers)):
backbone.layers[i].trainable = False
return backbone
# Creates the appearance embedding model generator
# Unused?
def get_branch_body(pool_size, downsample_rate, emb_len, use_roi_layer):
if use_roi_layer:
backbone_output = Input(shape=(None,None,1024//(32//downsample_rate)))
dummy_roi = Input(shape=(1, 4))
branch_body = RoiPoolingConv(pool_size=pool_size, num_rois=1)([backbone_output, dummy_roi]) # (None, 1, 7, 7, 512)
branch_body = Reshape((pool_size, pool_size, 1024//(32//downsample_rate)))(branch_body)
else:
roi_input = Input(shape=(pool_size, pool_size, 1024//(32//downsample_rate)))
branch_body = roi_input
branch_body = Lambda(lambda x: x)(branch_body)
branch_body = GlobalAveragePooling2D()(branch_body)
branch_body = Dense(emb_len)(branch_body)
branch_body = Lambda(lambda x: K.l2_normalize(x, axis=-1), name='normalize')(branch_body)
if use_roi_layer:
branch_body = Model([backbone_output, dummy_roi], branch_body, name='branch_body')
else:
branch_body = Model(roi_input, branch_body, name='branch_body')
branch_body.summary()
return branch_body
# Unused?
def get_embs(backbone, branch_body, downsample_rate, use_roi_layer, pool_size, name):
inp_roi = Input(shape=(1, 4), name='roi_{}'.format(name))
if backbone is None:
if use_roi_layer:
inp_img = Input(shape=(None,None,1024//(32//downsample_rate)))
branch = inp_img
else:
inp_img = Input(shape=(pool_size, pool_size, 1024//(32//downsample_rate)))
branch = inp_img
else:
inp_img = Input(shape=(None,None,3), name='img_{}'.format(name))
branch = backbone(inp_img)
if use_roi_layer:
branch = branch_body([branch, inp_roi])
return branch, inp_img, inp_roi
else:
branch = branch_body(branch)
return branch, inp_img
# Unused?
def get_triplet_model(path_weights, downsample_rate, pool_size, emb_len, gap, use_backbone, use_roi_layer, branch_type, **kwargs):
if use_backbone: backbone = get_backbone(path_weights, downsample_rate)
else: backbone = None
branch_body = get_branch_body(pool_size, downsample_rate, emb_len, gap, use_roi_layer, branch_type)
branchs = [ get_embs(backbone, branch_body, downsample_rate, use_roi_layer, pool_size, name) for name in ['A', 'P', 'N']]
output = concatenate([ b[0] for b in branchs], axis=-1, name='list_of_embds')
if use_roi_layer:
tripletModel = Model(inputs=[ img for _,img,_ in branchs] + [ roi for _,_,roi in branchs], outputs=output)
else:
tripletModel = Model(inputs=[ b[1] for b in branchs], outputs=output)
return tripletModel
def triplet_loss(y_true, y_pred, alpha = .2):
total_lenght = y_pred.shape.as_list()[-1]
print(int(total_lenght*1/3), int(total_lenght*2/3), int(total_lenght*3/3))
anchor = y_pred[:,0:int(total_lenght*1/3)]
positive = y_pred[:,int(total_lenght*1/3):int(total_lenght*2/3)]
negative = y_pred[:,int(total_lenght*2/3):int(total_lenght*3/3)]
print('anchor', anchor, anchor.shape, positive.shape, negative.shape)
# distance between the anchor and the positive
pos_dist = K.sum(K.square(anchor-positive),axis=1)
# distance between the anchor and the negative
neg_dist = K.sum(K.square(anchor-negative),axis=1)
# compute loss
basic_loss = pos_dist-neg_dist+alpha
loss = K.maximum(basic_loss,0.0)
return loss
def load_branch_body(model_folder_roi):
# Load architecture
model_arch = json.load(open(model_folder_roi + '/architecture.json', 'r'))
branch_body = model_from_json(model_arch, {'RoiPoolingConv': RoiPoolingConv})
model_params = json.load(open(model_folder_roi+'train_params.json', 'r'))
downsample_rate = model_params['downsample_rate']
# Load weights
branch_body_weights = sorted(os.listdir(model_folder_roi + '/weights'), key=lambda x: float(x[:-3].split('-')[-1][8:]))
if len(branch_body_weights) == 0: print(' ** No weights found for branch_body {}'.format(model_folder_roi))
else:
branch_body_weights = model_folder_roi + '/weights/' + branch_body_weights[0]
branch_body.load_weights(branch_body_weights, by_name=True, skip_mismatch=False)
# Create branch
branch_body = branch_body.get_layer('branch_body')
branch_body.name = branch_body.name + '_model_{}'.format(model_folder_roi)
if type(branch_body.layers[1]) != RoiPoolingConv:
branch_tp = json.load(open(model_folder_roi + '/train_params.json', 'r'))
inp_bck = Input(shape=(None,None,32*downsample_rate))
inp_roi = Input(shape=(1, 4))
roi_layer = RoiPoolingConv(pool_size=branch_tp['pool_size'], num_rois=1)([inp_bck, inp_roi])
roi_layer = Reshape((branch_tp['pool_size'], branch_tp['pool_size'], 32*downsample_rate))(roi_layer)
branch_body = branch_body(roi_layer)
branch_body = Model([inp_bck, inp_roi], branch_body)
return branch_body
# Unused?
def load_set_of_branches(path_roi_models):
inp_bck = Input(shape=(None,None,32*downsample_rate))
inp_roi = Input(shape=(1, 4))
branches = []
for prm in tqdm(path_roi_models, file=sys.stdout):
branch_body = load_branch_body(prm)
branch_body = branch_body([inp_bck, inp_roi])
branches.append(branch_body)
branch_set_model = Model([inp_bck, inp_roi], branches)
return branch_set_model